A hybrid of singular value decomposition and regression analysis collaborative filtering with linear incremental update method

• Main Authors: Wijak Srisujjalertwaja, Dussadee Praserttitipong
• Format: Journal
• Published: 2018
• Subjects: Agricultural and Biological Sciences
• Online Access: https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=84944533521&origin=inward; http://cmuir.cmu.ac.th/jspui/handle/6653943832/44478
• Institution: Chiang Mai University

© 2005 - 2015 JATIT & LLS. All rights reserved. Collaborative filtering (CF) approach comprises of several well-known techniques which successful in creating personalized recommendations. Singular Value Decomposition (SVD) based technique is the dominant class of CF techniques. The techniques rooted from SVD concept mostly return the high accuracy recommendation results than others. These SVD-based techniques are come up with the concept of model-based CF techniques, in which the relationship of historical ratings between users and items are learned. The learning processes usually perform in off-line mode. The model parameters are assessed according to this off-line process for constructing the knowledge models which are further use in on-line recommendation environments. However, the accuracy of the SVD-based technique is reduced according to the increasing number of solutions that they have to evaluate compared with the steady number of knowledge that they have learned, namely users-items sparse problem. On the other hand, the memorybased CF techniques are also suggested in literatures. These techniques are relied on collecting knowledge about ratings between users and items in computational area and re-computing the entire knowledge every time when the recommendations are called for. Thus, the users-items sparse problem is not an obstructer for memory-based CF techniques, because the new knowledge is always encompassed to the fundamental knowledge. Even though, memory-based CF techniques do not cause of difficulty in the users-items sparse problem, they impractical for implementing in on-line environment. Because of memory-based CF techniques take a lot of time for estimating just only one recommendation result. Hence, this paper proposed a hybrid of SVD-based technique and memory-based technique for CF. A regression analysis algorithm is proposed as a memory-based CF technique. The incremental update method with linear time refreshment also presented in this paper for making the practical on-line knowledge maintenance. The empirical experiment was established. The accuracy results acquired from the hybrid between the revolutionary RSVD technique and the proposed linear regression analysis with incremental update method depicted the highest accuracy, especially in users-items sparse situations.